CN110046709A - A kind of multi-task learning model based on two-way LSTM - Google Patents
A kind of multi-task learning model based on two-way LSTM Download PDFInfo
- Publication number
- CN110046709A CN110046709A CN201910326878.7A CN201910326878A CN110046709A CN 110046709 A CN110046709 A CN 110046709A CN 201910326878 A CN201910326878 A CN 201910326878A CN 110046709 A CN110046709 A CN 110046709A
- Authority
- CN
- China
- Prior art keywords
- lstm
- network
- weight
- input
- output
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
- G06N20/20—Ensemble learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Abstract
Public affairs of the present invention are related to the deep learning field of artificial intelligence field, have opened a kind of multi-task learning model based on two-way LSTM, to be completed at the same time in natural language processing part-of-speech tagging, language block identification, the name tasks such as Entity recognition.Major programme includes S1, defines single length memory network (LSTM) neural network;S2, the application method for defining two-way LSTM be using a LSTM network to input data string from left to right be sequentially input to LSTM network (L2R), while exporting result and being LSTM network (R2L) is sequentially inputted from what the right side was turned left to input data string using a LSTM network, while exporting result and beingMerge the output of L2R and R2L network as a result, i.e.S4, merge the output that output layer result obtains each subtask of word level.The present invention is applicable not only to natural language learning field, all can be used in other field.
Description
Technical field
A kind of multi-task learning model based on two-way LSTM, belongs to the deep learning scope of artificial intelligence field.
Background technique
Deep learning is a part of the wider machine learning method indicated based on learning data.Deep learning frame
Structure, such as deep neural network, depth confidence network and recurrent neural network etc. have been applied to computer vision, speech recognition, from
Right Language Processing, audio identification, social networks filtering, machine translation, bioinformatics, drug design, medical image analysis etc.
Field.The model result as caused by deep learning frame can compare favourably with human expert, or even be better than people in some cases
Class expert.
Deep learning can be divided into two kinds as machine learning: supervised learning and unsupervised learning.In recent years, depth
Habit technology obtains development at full speed with the raising that computer calculates power.It is all achieved out in fields such as information identification, recommended engines
The application effect of color.Meanwhile abundant experimental results prove that deep learning model has good robustness and generalization.
Currently, traditional disaggregated model is all modeled with single model for single task role, in production environment, this is built
Mould mode, which need to expend a large amount of manpowers and calculate power, to be carried out similar task to repeat modeling.For complicated problem, conventional model
Resolving ideas is that challenge is decomposed into simple and mutually independent subproblem individually to model solution, then remerges knot
Fruit, to obtain the result of initial challenge.And many problems cannot be decomposed into independent son one by one in production environment
Problem, even if can decompose, there is also stronger correlations between each subproblem.
Summary of the invention
The problem of for above-mentioned production environment, the present invention provides a kind of multi-task learnings based on two-way LSTM
Model framework solves the problems, such as that once training can only generate a model reply single task role to traditional single task model, enhance
The performance of model between inter-related task, compares the prior art, and this patent provides a kind of new method realization multitasking.Together
When, model of the present invention can solve the problems, such as to include being not limited to multiple word level tasks, Sentence-level task in natural language processing,
Being used in mixed way for word level and Sentence-level task can also be coped with.
In order to achieve the above objectives, the present invention adopts the following technical scheme:
One, multiple word levels or multiple Sentence-level tasks are directed to
S1, to define single LSTM neural network configuration first as follows:
ft=σg(Wfxt+Ufht-1+bf)
it=σg(Wixt+Uiht-1+bi)
ot=σg(Woxt+Uoht-1+bo)
ct=ft·ct-1+it·σc(Wcxt+Ucht-1+bc)
ht=ot·σh(ct)
Wherein xtFor the input of t moment;W is the corresponding input weight of different outputs, i.e. WfF is exported for t momenttIt is corresponding
Input weight, WiI is exported for t momenttCorresponding input weight, WoO is exported for t momenttCorresponding input weight, WcFor t moment
Export ctCorresponding input weight;U is the corresponding output weight of different outputs, i.e. UfF is exported for t momenttCorresponding output power
Weight, UiI is exported for t momenttCorresponding output weight, UoO is exported for t momenttCorresponding output weight, UcC is exported for t momentt
Corresponding output weight;B is the corresponding biasing of different outputs;σ is ReLu activation primitive.
S2, the application method for defining two-way LSTM (Bi-LSTM) are to input data string using a LSTM network from a left side
That turns right is sequentially input to LSTM network (L2R), while exporting result and beingUsing a LSTM network to input data string
LSTM network (R2L) is sequentially inputted from what the right side was turned left, while exporting result and beingIts concrete operations are as follows:
S2.1, setting neural network input layer, and configure initial weight;
S2.2, setting neural network hidden layer, hidden layer is set as 1 layer or 2 layers of shot and long term memory network, and ties S2.1
In initial weight be trained, the hidden layer weight after being trained;
S2.3, setting output layer, in conjunction with the weight output model result of hidden layer in S4.2;
S2.4, by input data according to from left to right sequentially input LSTM model, and the step of carrying out S2.1 to S2.3,
Obtain output result
S2.5, input data is sequentially inputted into LSTM model, and the step of carrying out S2.1 to S2.3 according to what is turned left from the right side,
Obtain output result
S3, the output result composition Bi-LSTM model for merging L2R and R2L network, i.e.,Use Bi-
The meaning of LSTM network is that Bi-LSTM network can be from the positive and negative both direction study of input sentence to history and future
Information, to be conducive to improve the predictive ability to follow-up work.
S4, the input by the output of Bi-LSTM as junior subtask, the stage can link the full connection of any number
Neural network model, and the output of each subtask is passed in other subtasks.The meaning of the operation is, Bi-LSTM mind
The feature selecting and feature weight assignment automated through network with the output data that other subtasks couple model entirely, passes through
Training iteration achievees the purpose that automatic parameter adjusts.The operation has the feature selecting mistake manually participated in different from traditional sense
Journey, model can be found that data inner link by iteration, find artificial indetectable fine feature, significantly improve model
Energy.Simultaneously as input of the output of other subtasks as current subtask, and exist centainly between each subtask
Task dependencies, the operation can provide stronger feature reference for each subtask, increase the concertedness between task.Separately
Outside, due to the increase of feature quantity, the data noise of subtask model is also increased to a certain extent, can make arriving for training
Model has stronger generalization ability.
S4, merge the output that output layer result obtains each subtask of word level, it is available by taking three subtasks as an example
Following formula:
Wherein m is the number of different subtasks, and b is biasing, and softmax is softmax function.
Two, for existing simultaneously word level and Sentence-level task
S1, as multiple word levels or multiple Sentence-level tasks, define single LSTM neural network configuration first
It is as follows:
ft=σg(Wfxt+Ufht-1+bf)
it=σg(Wixt+Uiht-1+bi)
ot=σg(Woxt+Uoht-1+bo)
ct=ft·ct-1+it·σc(Wcxt+Ucht-1+bc)
ht=ot·σh(ct)
Wherein xtFor the input of t moment;W is the corresponding input weight of different outputs, i.e. WfF is exported for t momenttIt is corresponding
Input weight, WiI is exported for t momenttCorresponding input weight, WoO is exported for t momenttCorresponding input weight, WcIt is carved for t cuns
Export ctCorresponding input weight;U is the corresponding output weight of different outputs, i.e. UfF is exported for t momenttCorresponding output power
Weight, UiI is exported for t momenttCorresponding output weight, UoO is exported for t momenttCorresponding output weight, UcC is exported for t momentt
Corresponding output weight;B is the corresponding biasing of different outputs;σ is ReLu activation primitive.
S2, the application method for defining two-way LSTM (Bi-LSTM) are to input data string using a LSTM network from a left side
That turns right is sequentially input to LSTM network (L2R), while exporting result and beingUsing a LSTM network to input data string
LSTM network (R2L) is sequentially inputted from what the right side was turned left, while exporting result and beingIts concrete operations are as follows:
S2.1, setting neural network input layer simultaneously configure initial weight;
S2.2, setting neural network hidden layer, hidden layer is set as 1 layer or 2 layers of shot and long term memory network, and combines
Initial weight in S2.1 is trained, the hidden layer weight after being trained;
S2.3, setting output layer, in conjunction with the weight output model result of hidden layer in S2.2;
S2.4, by input data according to from left to right sequentially input LSTM model, and the step of carrying out S2.1 to S2.3,
Obtain output result
S2.5, input data is sequentially inputted into LSTM model, and the step of carrying out S2.1 to S2.3 according to what is turned left from the right side,
Obtain output result
S3, the output result composition Bi-LSTM model for merging L2R and R2L network, i.e.,
S4, merge the output that output layer result obtains each subtask of word level, it is available by taking three subtasks as an example
Following formula:
S5, the input for constructing each subtask of Sentence-level are as follows:
xt=Ht+ym+ym+1+ym+2
S6, the input with the input constructed in S5 as next stage task, repeating S1 to S4 can be obtained next stage task
Output.
The present invention because using above technical scheme therefore have it is following the utility model has the advantages that
One, this patent requires to have used two LSTM models simultaneously in preposition model, one of LSTM model identification
The forward sequence (inputting from left to right) of input data, another LSTM model identify the reverse order of input data (from dextrad
Left input), which can allow neural network that can not only adequately learn the contextual information to historical information, while also can
Learn the contextual information to Future Information.
Two, the model combined strategy that similar Boosting has been used in preposition model, by two input datas not Tongfang
To LSTM neural network be combined, the input results as mid-module.
Three, this patent equally has originality in the internal structure of LSTM network, and this patent is in order to simplify LSTM network
Structure is all made of ReLu activation primitive as activation primitive on activation primitive.Meanwhile the operation can reach in calculating speed
To the promotion of calculating speed.
Four, model described in this patent has stronger expansion, specific manifestation are as follows:
4.1, when model structure is two-way LSTM model+multi task model, model can be completed at the same time multi task model
Corresponding multiple word level tasks.
4.2, when model structure is two-way LSTM model+multi task model+two-way LSTM model+multi task model, mould
Type can be completed at the same time word level task and Sentence-level task.
4.3, has the shared characteristic of weight between multi task model, i.e., the output weight of first sub- task model is with one
Certainty ratio shares to second sub- task model, and so on.Meanwhile the model output of first subtask is used as second son
Task model enters modular character.The operation provides the data characteristics of more various dimensions for subsequent child task model, which can be with
The boosting being analogous in integrated study improves the predictive ability of model entirety, simultaneously because subtask model is specific
Task has a different, which can be considered as the input of higher level subtask model to a certain extent under noise data is
Grade subtask model provides data noise, the generalization ability of enhancer task model.
4.5, this patent two, for existing simultaneously word level and the meaning of S6 in Sentence-level task is:
4.5.1, as the defeated of the Bi-LSTM of next multitask module after the output of higher level's multitask module being integrated
Enter.The stage, link of the Bi-LSTM as two multitask modules re-start feature sampling to input data, and will sampling
As a result it is passed in each subtask model of next multitask module.The meaning of the operation is that Bi-LSTM neural network is to higher level
The output data that the Bi-LSTM model row output of multitask module couples model with other subtasks entirely re-starts automation
Feature selecting and feature weight assignment achieve the purpose that automatic parameter adjusts by training iteration.The operation is anticipated different from tradition
Justice has the feature selection process manually participated in, and model can be found that data inner link by iteration, finds and manually is not easy to send out
Existing fine feature, significantly improves model performance.
4.5.2, since junior's multitask module equally uses Bi-LSTM model as the connection of two multitask modules,
Subsequent multitask module can be made with reference also to the history of higher level's multitask module and following information, can dynamically adjust junior
The weight configuration of each model in multitask module.
4.5.3, the operation is also different from conventional multilayer single task Bi-LSTM model structure, due to higher level's multitask module
Each subtask input of the output as junior's multitask module, and because of there are certain tasks between each subtask
Correlation, the operation can provide stronger feature reference for junior's multitask module, increase the concertedness between subtask.Separately
Outside, due to the increase of feature quantity, also increase the data noise of junior's multi task model to a certain extent, arriving for training can be made
Model have stronger generalization ability.
4.5.4, compared with this patent model, conventional double single task Bi-LSTM model lacks intermediate each subtask mould
Block, thus will lead to the second level Bi-LSTM model of conventional double single task Bi-LSTM to the output of higher level's Bi-LSTM model into
Row excessively sampling (because there are strong correlations with prediction result for the output of higher level Bi-LSTM model), so as to cause traditional double
Layer single task Bi-LSTM model is poor to the Generalization Capability of task.And this patent model is shown by constructing each subtask module
Ground introduces the input feature vector to the unobvious relevant data of prediction result as second level Bi-LSTM model, can significantly avoid
The problem of appeared in conventional double single task Bi-LSTM, improve the generalization ability of model.
4.5, comparing patent document, " CN109375776A- acts intention assessment based on the EEG signals of multitask RNN model
Method " uses model only as more classification tasks in the multitask stage, and this patent requires model to can be used not only for more classification times
It is engaged in (word level task), while can be used on overstepping one's bounds generic task, such as the emotion recognition of text, text output prediction, spelling is entangled
The tasks such as just.
Detailed description of the invention
Fig. 1 is LSTM Construction of A Model and internal structure chart used in the present invention;
Fig. 2 is word level/Sentence-level model structure of multitask in the present invention;
Fig. 3 is the word level and Sentence-level model structure of multitask in the present invention.
Specific embodiment
Model is specifically described below in conjunction with attached drawing:
The present invention provides a kind of multi-task learning models based on two-way LSTM, it is characterised in that following three points:
S1.1, a unified multi-task learning model is proposed, wherein including minimal number of RNN (circulation nerve net
Network) (two layers/mono- layer) of the layer natural language processing task for word level and Sentence-level.
S1.2, the model, which can be used for multiple word level tasks (such as POS, Chunk, NER), can be used for multiple Sentence-levels
Task (such as sentiment analysis).It is also possible to learn the task of both types together.
S1.3, the number of plies for keeping less RNN/LSTM neural network change output layer only to accelerate training speed.
The neural network number of plies that model uses in S1.1 is less, to cope with different types of natural language processing task.
It can accelerate the speed of model training using the less neural network number of plies.
Model is different from a model in traditional single task model and is only capable of completing a task, patent requirements 1 in S1.2
The model can carry out unified study to the task of word level in natural language processing or Sentence-level simultaneously.Learn due to unified
Task have correlation, some factors can be shared between multiple tasks and (including are not limited into modular character, the model of neural network
Parameter etc.), to improve the effect and Generalization Capability of each task in model.
Embodiment
One, multiple word levels or multiple Sentence-level tasks are directed to
S1, as shown in Figure 1, to define single LSTM neural network configuration first as follows:
ft=σg(Wfxt+Ufht-1+bf)
it=σg(Wixt+Uiht-1+bi)
ot=σg(Woxt+Uoht-1+bo)
ct=ft·ct-1+it·σc(Wcxt+Ucht-1+bc)
ht=ot·σh(ct)
S2, the application method for defining two-way LSTM (Bi-LSTM) are to input data string using a LSTM network from a left side
That turns right is sequentially input to LSTM network (L2R), while exporting result and beingUsing a LSTM network to input data string
LSTM network (R2L) is sequentially inputted from what the right side was turned left, while exporting result and being
S3, merge the output of L2R and R2L network as a result, i.e.The part as shown in "+" in Fig. 2;
S4, merge the output that output layer result obtains each subtask of word level, it is available by taking three subtasks as an example
Following formula:
Two, for existing simultaneously word level and Sentence-level task
S1, as multiple word levels or multiple Sentence-level tasks, define single LSTM neural network structure first
(as shown in Figure 1) is as follows:
ft=σg(Wfxt+Ufht-1+bf)
it=σg(Wixt+Uiht-1+bi)
ot=σg(Woxt+Uoht-1+bo)
ct=ft·ct-1+it·σc(Wcxt+Ucht-1+bc)
ht=ot·σh(ct)
S2, the application method for defining two-way LSTM (Bi-LSTM) are to input data string using a LSTM network from a left side
That turns right is sequentially input to LSTM network (L2R), while exporting result and beingUsing a LSTM network to input data string
LSTM network (R2L) is sequentially inputted from what the right side was turned left, while exporting result and being
S3, merge the output of L2R and R2L network as a result, i.e.The part as shown in the "+" of the lower part Fig. 3;
S4, the output for merging each subtask that output layer result obtains word level are available by taking three subtasks as an example
Following formula:
S5, the part as shown in the "+" of the top Fig. 3, the input for constructing each subtask of Sentence-level are as follows:
x′t=Ht+ym+ym+1+ym+2
S6, the input with the input constructed in S5 as next stage task, repeating S1 to S4 can be obtained next stage task
Output.
Claims (3)
1. a kind of multi-task learning model based on two-way LSTM, which comprises the following steps:
S1, single LSTM neural network is defined;
S2, the application method for defining two-way LSTM be using a LSTM network to input data string sequentially inputting from left to right
To LSTM network L2R, while exporting result and beingThe sequence turned left to input data string from the right side using a LSTM network is defeated
Enter LSTM network R2L, while exporting result and being
S3, merge the output of LSTM network L2R and LSTM network R2L network as a result, i.e.
S4, merge the output that output layer result obtains each subtask of word level, it is available as follows by taking three subtasks as an example
Formula:
Wherein m is the number of different subtasks, and b is biasing, and softmax is softmax function;
S5, the input for constructing each subtask of Sentence-level are as follows:
x′t=Ht+ym+ym+1+ym+2
S6, the input with the input constructed in S5 as next stage task, repeating S1 to S4 can be obtained the defeated of next stage task
Out;
For multiple word levels or multiple Sentence-level tasks, step S1-S4 is executed, it is in charge of a grade with sentence for word level is existed simultaneously
Business executes step S1-S6.
2. a kind of multi-task learning model based on two-way LSTM according to patent requirements 1, which is characterized in that single LSTM
Neural network configuration is as follows:
ft=σg(Wfxt+Ufht-1+bf)
it=σg(Wixt+Uiht-1+bi)
ot=σg(Woxt+Uoht-1+bo)
ct=ft·ct-1+it·σc(Wcxt+Ucht-1+bc)
ht=ot·σh(ct)
Wherein xtFor the input of t moment;W is the corresponding input weight of different outputs, i.e. WfF is exported for t momenttCorresponding input
Weight, WiI is exported for t momenttCorresponding input weight, WoO is exported for t momenttCorresponding input weight, WcFor t moment output
ctCorresponding input weight;U is the corresponding output weight of different outputs, i.e. UfF is exported for t momenttCorresponding output weight, Ui
I is exported for t momenttCorresponding output weight, UoO is exported for t momenttCorresponding output weight, UcC is exported for t momenttIt is corresponding
Export weight;B is the corresponding biasing of different outputs;σ is ReLu activation primitive.
3. a kind of multi-task learning model based on two-way LSTM according to patent requirements 1, which is characterized in that step S2 packet
Include following steps:
S2.1, setting neural network input layer, and configure initial weight;
S2.2, setting neural network hidden layer, hidden layer is set as 1 layer or 2 layers of shot and long term memory network, and combines in S2.1
Initial weight be trained, the hidden layer weight after being trained;
S2.3, setting output layer, in conjunction with the weight output model result of hidden layer in S2.2;
S2.4, by input data according to from left to right sequentially input LSTM model, and the step of carrying out S2.1 to S2.3, obtain
Export result
S2.5, input data is sequentially inputted into LSTM model, and the step of carrying out S2.1 to S2.3 according to what is turned left from the right side, obtained
Export result
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910326878.7A CN110046709A (en) | 2019-04-22 | 2019-04-22 | A kind of multi-task learning model based on two-way LSTM |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910326878.7A CN110046709A (en) | 2019-04-22 | 2019-04-22 | A kind of multi-task learning model based on two-way LSTM |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110046709A true CN110046709A (en) | 2019-07-23 |
Family
ID=67278590
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910326878.7A Pending CN110046709A (en) | 2019-04-22 | 2019-04-22 | A kind of multi-task learning model based on two-way LSTM |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110046709A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111209738A (en) * | 2019-12-31 | 2020-05-29 | 浙江大学 | Multi-task named entity recognition method combining text classification |
CN112149119A (en) * | 2020-09-27 | 2020-12-29 | 苏州遐视智能科技有限公司 | Dynamic active security defense method and system for artificial intelligence system and storage medium |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108229582A (en) * | 2018-02-01 | 2018-06-29 | 浙江大学 | Entity recognition dual training method is named in a kind of multitask towards medical domain |
CN109375776A (en) * | 2018-10-30 | 2019-02-22 | 东北师范大学 | EEG signals based on multitask RNN model act intension recognizing method |
CN109460466A (en) * | 2018-09-20 | 2019-03-12 | 电子科技大学 | It is a kind of based on relationship analysis method between the two-way length of the multitask in short-term implicit sentence of memory network |
-
2019
- 2019-04-22 CN CN201910326878.7A patent/CN110046709A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108229582A (en) * | 2018-02-01 | 2018-06-29 | 浙江大学 | Entity recognition dual training method is named in a kind of multitask towards medical domain |
CN109460466A (en) * | 2018-09-20 | 2019-03-12 | 电子科技大学 | It is a kind of based on relationship analysis method between the two-way length of the multitask in short-term implicit sentence of memory network |
CN109375776A (en) * | 2018-10-30 | 2019-02-22 | 东北师范大学 | EEG signals based on multitask RNN model act intension recognizing method |
Non-Patent Citations (2)
Title |
---|
廖祥文等: "《基于多任务迭代学习的论辩挖掘方法》", 《计算机学报》 * |
郭江: "《基于分布表示的跨语言跨任务自然语言分析》", 《中国博士学位论文全文数据库 信息科技辑》 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111209738A (en) * | 2019-12-31 | 2020-05-29 | 浙江大学 | Multi-task named entity recognition method combining text classification |
CN112149119A (en) * | 2020-09-27 | 2020-12-29 | 苏州遐视智能科技有限公司 | Dynamic active security defense method and system for artificial intelligence system and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US11514305B1 (en) | Intelligent control with hierarchical stacked neural networks | |
US20230419074A1 (en) | Methods and systems for neural and cognitive processing | |
CN106560848B (en) | Novel neural network model for simulating biological bidirectional cognitive ability and training method | |
Natesan Ramamurthy et al. | Model agnostic multilevel explanations | |
CN106503654A (en) | A kind of face emotion identification method based on the sparse autoencoder network of depth | |
Lin et al. | Attribute-Aware Convolutional Neural Networks for Facial Beauty Prediction. | |
CN104732274A (en) | Intelligent computer | |
CN107451230A (en) | A kind of answering method and question answering system | |
CN108427665A (en) | A kind of text automatic generation method based on LSTM type RNN models | |
CN108829737B (en) | Text cross combination classification method based on bidirectional long-short term memory network | |
CN113987179A (en) | Knowledge enhancement and backtracking loss-based conversational emotion recognition network model, construction method, electronic device and storage medium | |
CN113254675B (en) | Knowledge graph construction method based on self-adaptive few-sample relation extraction | |
Tirumala | Evolving deep neural networks using coevolutionary algorithms with multi-population strategy | |
Wu et al. | Optimized deep learning framework for water distribution data-driven modeling | |
CN110046709A (en) | A kind of multi-task learning model based on two-way LSTM | |
Hebbar | Augmented intelligence: Enhancing human capabilities | |
Monroe | Neurosymbolic ai | |
CN112000793B (en) | Man-machine interaction oriented dialogue target planning method | |
CN110297894A (en) | A kind of Intelligent dialogue generation method based on auxiliary network | |
CN112560440A (en) | Deep learning-based syntax dependence method for aspect-level emotion analysis | |
CN116975743A (en) | Industry information classification method, device, computer equipment and storage medium | |
CN111753995A (en) | Local interpretable method based on gradient lifting tree | |
Goertzel et al. | The Novamente artificial intelligence engine | |
Li et al. | Multimodal information-based broad and deep learning model for emotion understanding | |
CN114169408A (en) | Emotion classification method based on multi-mode attention mechanism |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190723 |
|
RJ01 | Rejection of invention patent application after publication |